Impacts of historical records on extreme flood variations over the conterminous United States

Evaluation of flood variations over time, especially for floods with large return periods, is of great significance to flood risk assessment. ‘Historical’ data that have been recorded before the construction of a gauging station provide an effective way to analyse the temporal changes of extreme floods. Here, comparison of maximum likelihood method, L-moment method and Bayesian theory are made to calculate the Gumbel distribution parameters via Monte Carlo simulation experiment. The best option is applied to 37 unregulated rivers over the conterminous United States to analyse their 100-year flood variations. The Monte Carlo simulation results indicate that L-moment method is substantially better than the other two estimators for both systematic and unsystematic series. Over 70% of studied river catchments detect 100-year flood decrease when the historical data are considered. The impacts of historical records on 100-year flood variation estimations are closely related to censoring threshold and historical period length.